DETECTION OF FADING-IN/OUT TEXT REGIONS FROM MPEG VIDEOS BASED ON A MARKOV MODEL

摘要

This paper proposes a new detection technique for fading in/out text regions from MPEG1 bitstreams based on a Markov model. We focus on the properties that the luminance of a text region varies linearly in fading, and that text regions require larger amount of bits than non-text regions. These features are extracted directly from a target MPEG1 bitstream and they are fed into the Markov model to detect the status of text regions. The advantage of our approach lies in the fact that erroneous feature values do not directly affect the resulting text region detection due to a good behavior of the Markov model. A recall of 80% and a precision of up to 90% have been obtained for several sample video sequences.